TL;DR
- MOE 2.8T total and 104B active parameters, 1M context window
- it's probably the same architecture as K3 (Kimi Delta Attention plus gated MLA, with native MXFP4 weights).
- training: performance gains are from large-scale RL across hundreds of domains, with the focus areas in full-project game-engine coding, 3D and CAD work in Blender, and finance.
- coding: 65.7% on Terminal-Bench 4.0, against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. leads SWE Marathon with 51.7, ahead of Opus 5.5 at 46.3.
- agentic: top of the chart on OfficeQA Pro v2 (64.4), Agents' Last Exam (43.2), HLE with tools (70.0) and BrowseComp (92.1). falls well behind on Terminal-Bench Science (38.1 vs 68.1), ProgramBench, and HLE without tools.
- against open peers: beats DeepSeek-V4.1-Flash and GLM-5.3 on every chart in the summary image, but DeepSeek beats it on DeepSWE (77.0 vs 75.5).
- hallucination: on Moonshot's internal test, the hallucination rate fell from 23.4% on K3 to 8.3%. GPT-6 Astra scores 5.5 on the same test.
- cyber safety: cyber-related data was removed from training for the open-weight release. That lowers its score on Moonshot's exploit test from 90.5% to 81.0%, which is still far above K3's 33.3%.
- price: $2 input and $10 output per 1M tokens, with cached input at $0.20. two-thirds of K3's price.
- open weights by Nov 4, 2026